A lot of the more cynical commenters here are misunderstanding the breakthroughs that have lead to the novel discovery, and underpin the emerging ML revolution

A lot of the more cynical commenters here are misunderstanding the breakthroughs that have lead to the novel discovery, and underpin the emerging ML revolution that we are just beginning to witness. Yes, partly the results of this study are due to increased compute, but I'd say that's only about 50% of the secret sauce. The other 50% is attributable to many very recent developments in the field ML which are gradually coming together in solutions to a range of problems - deep learning, convolutional networks, new architectures (GAN, transformer, RNN, LSTM, autoencoders, etc), regularization techniques, gradient control, hyperparameter optimization...the list is quite long, and every SOTA neural network inevitably incorporates a large proportion of these small steps towards the massive leaps that are being taken in the applied ML space. The concepts of perceptrons and gradient descent may be 50+ years old, but an entire body of theory has been explosively developed in the last decade, such that comparing modern ML to what was being done just a decade ago is akin to the difference between, say, programming theory now and 100 years ago. And for the same reason - just as compute advances enabled development of computer science, so to have recent hardware advances unlocked a new level of machine learning. But I'd like to point out that, even moreso than software development, very little of the grand breakthroughs we will soon see will be possible without multidisciplinary domain knowledge. It is very difficult to effectively apply ML without a solid technical understanding of the properties of the applied data space, which for real world applications are constrained by physical laws and represented and communicated best by mathematical descriptions. ML engineering is a generalist's game - and what we are going to find is that the most successful ML engineers come from broadly applicable, math heavy backgrounds - physics in particular, electrical engineering, to a lesser degree mathematics, etc - because ultimately training a neural network comes down to adequately sampling a problem space and curating data with an intuition which is most ideally developed by the study of mathematics. It is a very general view of the world which is difficult to communicate to someone who is not experienced with higher math. The current wave of applied ML startups will see a high rate of failure - because ML is still being treated as an extension of programming, in the sense that you expect to be able to hire a bunch of pure developers to translate a specialist's knowledge into code. But this emerging field is different, the few startups that succeed in the applied ML space will be those that are able to find the rare domain experts who have picked up ML along with their math and science experience. There will effectively emerge two classes of ML engineers with substantially different levels of compensation - the coveted generalists who have cross-pollinated with math heavy disciplines, and the rest.